All Times EDT
Keywords: Biostatistics, Machine Learning, Real Time, Big Data, Data Science, Random Forest, Clinical Trial
Controlling blood glucose is essential to the health and well-being of patients with diabetes; patient outcomes are affected by absolute glucose levels and measurement variability. Continuous Glucose Monitoring (CGM) captures real-time measurements every 5- 15 minutes, 24 hours a day. Traditionally, such data has been analyzed as weekly averages. Averaging loses much of the power from this rich data by ignoring differences in activities of daily living and their potential impact on glucose, attributing all deviations from the mean to random variability. We used various methods from statistics and machine learning to examine factors that may account for variability in the CGM data and to predict changes in blood glucose, such as gradient boosting machines. We also employed virtual twins and random forests to investigate important baseline characteristics and explore subgroups with potential for treatment benefit. We considered potential new endpoints and analyses available from real-time data that may be more meaningful for patients and physicians. Finally, we hypothesized that the first 2 weeks of CGM data could be used to forecast glucose levels in subsequent weeks and validated this hypothesis using an expanded dataset with three additional trials.